基于bp-MRI和VB-Net的临床显著性前列腺癌自动分割和机器学习分类的研究  

A Study of Automatic Segmentation and Machine Learning Classification of Clinically Significant Prostate Cancer Based on bp-MRI and VB-Net

作　　者：彭涛[1] 肖建明[1] 李林 蒲冰洁 夏朝阳 牛翔科 王宗勇[1] 曾小辉[1] 陈林[2] 杨进[2] 李佽[3] PENG Tao;XIAO Jianming;LI Lin(Department of Radiology,the Affiliated Hospital of Chengdu University,Chendu,Sichuan Province 610081,P.R.China)

机构地区：[1]成都大学附属医院放射科,610081 [2]成都大学附属医院泌尿外科,610081 [3]成都大学附属医院病理科,610081 [4]上海联影智能医疗科技有限公司,201807

出　　处：《临床放射学杂志》2022年第3期519-524,共6页Journal of Clinical Radiology

基　　金：四川省卫生健康委员会科研课题项目(编号:18PJ150);成都市卫生健康委员会医学科研课题项目(编号:2018055、2020177、2021036)。

摘　　要：目的评价基于双参数MRI的深度学习自动分割与机器学习分类模型,探索其在临床显著性前列腺癌(CSPC)诊断中的应用。方法纳入409例前列腺患者MRI检查资料,在DWI、ADC和T;WI中应用VB-Net模型分别进行病灶自动分割和腺体自动分割,生成感兴趣区(ROI),病灶自动分割时将分割阈值设置为不同数值分别重复进行。分别提取病灶ROI和腺体ROI中的纹理特征,进行Lasso特征选择,建立、训练随机森林、支持向量机和Logistic回归模型并进行验证。结果病灶分割中分割阈值分别为0.9、0.5、0.1时,假阴性率分别为0.462、0.273、0.182,假阳性率分别为0.134、0.419、0.661;当分割阈值设为0.5,病灶自动分割后进行纹理分析和机器学习分类,3种模型ROC曲线的AUC为0.76~0.792;腺体分割后进行纹理分析和机器学习分类,3种模型ROC曲线的AUC为0.827~0.855。结论采用基于前列腺bp-MRI的VB-Net模型对CSPC病灶具有一定的自动分割、分类能力,结合进一步的机器学习能较好地诊断CSPC;VB-Net模型对腺体自动分割后再进行机器学习分类,也对CSPC有较好的诊断价值。在提高全自动病灶自动分割能力方面还需要进一步的深入研究。Objective To evaluate the deep learning automatic segmentation and machine learning classification models based on biparametric Magnetic Resonance Imaging(bp-MRI),and to research its application in the diagnosis of clinically significant prostate cancer.Methods Data of 409 patients with prostate MRI were included.VB-Net model was used to perform automatic segmentation of lesions and glands on DWI,ADC and T;WI images,respectively,to generate regions of interest(ROI).The segmentation thresholds were set to different values for automatic segmentation of lesions and repeated respectively.Texture features in lesion ROI and gland ROI were extracted respectively.After texture feature extraction,feature selection was carried out by Lasso method.Random forest,support vector machine and Logistic regression models were established for training and verification.Results When the segmentation thresholds were 0.9,0.5 and 0.1,the false negative rates were 0.462,0.273 and 0.182,and the false positive rates were 0.134,0.419 and 0.661,respectively.When the segmentation threshold was set to 0.5,texture analysis and machine learning classification were performed after automatic segmentation of lesions,the AUC of ROC curves of the three models ranged from 0.76-0.792.After gland segmentation,texture analysis and machine learning classification were performed,and the AUC of the ROC curves of the three models ranged from 0.827 to 0.855.Conclusion VB-Net model based on prostate bp-MRI has certain ability of automatic segmentation and classification of CSPC lesions.Combined with further machine learning,it can better diagnose CSPC.The method of automatic gland segmentation by VB-Net model followed by machine learning classification also has better diagnostic value for CSPC.Further research is needed to improve the ability of automatic lesion segmentation.

关 键 词：前列腺癌 双参数磁共振成像 深度学习 卷积神经网络 纹理分析 机器学习 

分 类 号：R737.25[医药卫生—肿瘤]